Argthtjtr

I've been designing my own handheld gaming device based around an AVR microcontroller and a small OLED display.

I started off with a monochrome display 128x64 pixels and can comfortably draw to it at over 60 frames per second.

I recently reworked it to use an RGB OLED, 128x128 pixels without really thinking too much only to find I could only achieve about 4 FPS. After some thought and careful refactoring I can get that up to ~12fps if I don't care too much about doing anything else!

My question is - how did a device like the GBA (Game Boy Advance) achieve a frame rate of nearly 60fps? I thought about having a separate 'graphics processor' but realised I would still be bottlenecked transferring the display data to that.

I also wondered about using the vestigial 8-bit parallel interface most of these screens tend to have, which might net me an 8x speed up, except that modern MCUs don't tend to have hardware parallel interfaces like they do for serial and bit-banging will likely eat up a lot of the speed gain.

What other options exist?

I'm currently using an ATmega1284P connected to an SSD1306 OLED controller via USART-SPI. That's the monochrome version.

The colour screen was an SSD1351, not originally connected to hardware SPI. I wasn't convinced it would make enough difference, it's just too slow overall

I know I can get faster MCUs, but I want to know what other options I could explore - the GBA processor is much slower than my 1284!

Other answers cover your question pretty well at an abstract level (hardware), but having actual experience with the GBA in particular I figured a more detailed explanation may be worth while.

The GBA had many drawing modes and settings which could be used to control how the graphics processor interpreted the video RAM, but one thing was inescapable: the frame rate. The graphic processor was drawing to the screen in a nearly (more on this below) constant loop. (This is likely the most relevant bit for your question.)

It would draw one line at a time taking a very short break between each. After drawing the last line for the frame it would take a break roughly equal to the time it takes to draw 30 lines. Then start again. The timing of each line, and the timing of each frame were all predetermined and set in stone. In a lot of ways the graphics processor was really the master of that system and you needed to write your games around its behavior, because it would continue doing what it did whether you were ready or not.

Roughly 75-80% of the time it was actively pushing to the screen. What frame rates could you accomplish if you were doing the same?

That 80% of the time was also what the CPU had to process user input, calculate game state, and load sprites/tiles to areas of VRAM that were currently off screen (or at least not included in the current line being drawn).

The 20% between frames, was all the CPU had to tweak video settings or RAM that would impact the whole next frame.

At the end of each line, the graphics processor would send a line sync interrupt to the CPU. This interrupt could be used to tweak settings on a few sprites, or a few background layers (this is how you can get an effect like a conical spotlight, by changing the size and location of one of the rectangular masks between each line drawn. As far as the hardware is concerned all those regions are rectangular.). You have to be careful to keep these updates small and finish before the graphic processor starts drawing the next line or you can get ugly results. Any time spent processing these interrupts also cut into that 80% of the CPU's processing time...

For games that got the most out of this system, neither the CPU nor the graphic processor ever took a real break; each were chasing the other around the loop updating what the other wasn't currently looking at.

The key feature of all the games consoles that distinguished them from early PCs and virtually all home computers(1) was hardware sprites.

The linked GBA programming guide shows how they work from the main processor point of view. Bitmaps representing player, background, enemies etc are loaded into one area of memory. Another area of memory specifies the location of the sprites. So instead of having to re-write all of video RAM every frame, which takes a lot of instructions, the processor just has to update the location of the sprites.

The video processor can then work pixel by pixel to determine which sprite to draw at that point.

However, this requires dual-port RAM shared between the two, and I think in the GBA the video processor is on the same chip as the main ARM and secondary Z80 processor.

(1) Notable exception: Amiga

"My question is - how did a device like the GBA achieve a frame rate of nearly 60fps?"

To answer just the question, they did it with a graphics processer. I'm pretty sure the Game Boy used sprite graphics. At a top level, that means that the graphics processor gets loaded things like an image of a background, and an image of Mario, and an image of Princess Peach, etc. Then the main processor issues commands like "show the background offset by this much in x and y, overlay Mario image #3 at this x, y position", etc. So the main processor is absolutely positively not concerned with drawing each pixel, and the graphics processor is absolutely positively not concerned with computing the state of the game. Each is optimized for what it needs to do, and the result is a pretty good video game without using a lot of computation power.

The GBA had a pretty slow processor. The ARM7 is very nice; they just ran it slow and gave it next to no resources.

There is a reason why a lot of Nintendo games at that point and before were side-scrollers. HARDWARE. It is all done in hardware. You had multiple layers of tiles plus one or more sprites and the hardware did all the work to extract pixels from those tables and drive the display.

You build the tile set up front and then had a smallish memory that was a tile map. Want the lower left tile to be tile 7? You put a 7 in that memory location. Want the next tile over to be tile 19? In the tile set, you put a 19 there, and so on for each layer that you have enabled. For the sprite, you simply set the x/y address. You can also do scaling and rotation by setting some registers and the hardware takes care of the rest.

Mode 7, if I remember right, was a pixel mode, but that was like a traditional video card where you put bytes in that cover the color for a pixel and the hardware takes care of the video refresh. I think you could ping pong or at least when you had a new frame you could flip them, but I don't remember right. Again, the processor was fairly underclocked for that day and age and didn't have too many fast resources. So while some games were mode 7, a lot were tile based side-scrollers...

If you want a solution that is a high frame rate, you need to design that solution. You can't just take any old display you find and talk to it via SPI or I²C or something like that. Put at least one framebuffer in front of it, ideally two, and have row and column control if possible over that display.

A number of the displays I suspect you are buying have a controller on them that you are actually talking to. If you want GBA/console type performance you create/implement the controller. Or you buy/build with a GPU/video chip/logic blob, and use HDMI or other common interface into a stock monitor.

Just because a bicycle has tires and a chain and gears doesn't mean it can go as fast as a motorcycle. You need to design the system to meet your performance needs, end to end. You can put that bicycle wheel on that motorcycle, but it won't perform as desired; all of the components have to be part of the overall design.

Asteroids worked this way too; it only needed one 6502. The vector graphics were done with separate logic; the 6502 sent a tiny string of data to the vector graphics controller, which used a ROM and that data to do the xy plotting of the beam and z, on/off... Some standups had separate processors to handle audio and video separate from the processor computing the game. Of course today the video is handled by some hundreds, if not thousands, of processors that are separate from the main processor...